3 files changed, 81 insertions, 24 deletions

diff --git a/indra/newview/lldrawpoolavatar.cpp b/indra/newview/lldrawpoolavatar.cpp
index 2cd9fef9f5..b509b363b7 100644
--- a/indra/newview/lldrawpoolavatar.cpp
+++ b/indra/newview/lldrawpoolavatar.cpp
@@ -123,7 +123,7 @@ void LLDrawPoolAvatar::prerender()
 	
 	if (sShaderLevel > 0)
 	{
-		sBufferUsage = GL_STATIC_DRAW_ARB;
+		sBufferUsage = GL_DYNAMIC_DRAW_ARB;
 	}
 	else
 	{
@@ -323,7 +323,7 @@ void LLDrawPoolAvatar::renderShadow(S32 pass)
 
 S32 LLDrawPoolAvatar::getNumPasses()
 {
-	return LLPipeline::sImpostorRender ? 1 : 3;
+	return LLPipeline::sImpostorRender ? 1 : 4;
 }
 
 void LLDrawPoolAvatar::render(S32 pass)
@@ -586,6 +586,7 @@ void LLDrawPoolAvatar::beginRigged()
 void LLDrawPoolAvatar::endRigged()
 {
 	sVertexProgram = NULL;
+	LLVertexBuffer::unbind();
 	gSkinnedObjectSimpleProgram.unbind();
 	LLVertexBuffer::sWeight4Loc = -1;
 }
diff --git a/indra/newview/llviewerjointmesh.cpp b/indra/newview/llviewerjointmesh.cpp
index 92029d10f6..90a5a29bb4 100644
--- a/indra/newview/llviewerjointmesh.cpp
+++ b/indra/newview/llviewerjointmesh.cpp
@@ -661,6 +661,8 @@ void LLViewerJointMesh::updateFaceSizes(U32 &num_vertices, U32& num_indices, F32
 //-----------------------------------------------------------------------------
 // updateFaceData()
 //-----------------------------------------------------------------------------
+static LLFastTimer::DeclareTimer FTM_AVATAR_FACE("Avatar Face");
+
 void LLViewerJointMesh::updateFaceData(LLFace *face, F32 pixel_area, BOOL damp_wind)
 {
 	mFace = face;
@@ -670,6 +672,8 @@ void LLViewerJointMesh::updateFaceData(LLFace *face, F32 pixel_area, BOOL damp_w
 		return;
 	}
 
+	LLFastTimer t(FTM_AVATAR_FACE);
+
 	LLStrider<LLVector3> verticesp;
 	LLStrider<LLVector3> normalsp;
 	LLStrider<LLVector2> tex_coordsp;
@@ -688,30 +692,76 @@ void LLViewerJointMesh::updateFaceData(LLFace *face, F32 pixel_area, BOOL damp_w
 			face->mVertexBuffer->getIndexStrider(indicesp);
 			stop_glerror();
 
-			for (U16 i = 0; i < mMesh->getNumVertices(); i++)
+			verticesp += mMesh->mFaceVertexOffset;
+			tex_coordsp += mMesh->mFaceVertexOffset;
+			normalsp += mMesh->mFaceVertexOffset;
+			vertex_weightsp += mMesh->mFaceVertexOffset;
+			clothing_weightsp += mMesh->mFaceVertexOffset;
+
+			U32* __restrict v = (U32*) verticesp.get();
+			const U32 vert_skip = verticesp.getSkip()/sizeof(U32);
+
+			U32* __restrict tc = (U32*) tex_coordsp.get();
+			const U32 tc_skip = tex_coordsp.getSkip()/sizeof(U32);
+
+			U32* __restrict n = (U32*) normalsp.get();
+			const U32 n_skip = normalsp.getSkip()/sizeof(U32);
+			
+			U32* __restrict vw = (U32*) vertex_weightsp.get();
+			const U32 vw_skip = vertex_weightsp.getSkip()/sizeof(U32);
+
+
+			U32* __restrict cw = (U32*) clothing_weightsp.get();
+			const U32 cw_skip = vertex_weightsp.getSkip()/sizeof(U32);
+
+			const U32* __restrict coords = (U32*) mMesh->getCoords();
+			const U32* __restrict tex_coords = (U32*) mMesh->getTexCoords();
+			const U32* __restrict normals = (U32*) mMesh->getNormals();
+			const U32* __restrict weights = (U32*) mMesh->getWeights();
+			const U32* __restrict cloth_weights = (U32*) mMesh->getClothingWeights();
+
+			const U32 num_verts = mMesh->getNumVertices();
+
+			U32 i = 0;
+			do
 			{
-				verticesp[mMesh->mFaceVertexOffset + i] = *(mMesh->getCoords() + i);
-				tex_coordsp[mMesh->mFaceVertexOffset + i] = *(mMesh->getTexCoords() + i);
-				normalsp[mMesh->mFaceVertexOffset + i] = *(mMesh->getNormals() + i);
-				vertex_weightsp[mMesh->mFaceVertexOffset + i] = *(mMesh->getWeights() + i);
-				if (damp_wind)
-				{
-					clothing_weightsp[mMesh->mFaceVertexOffset + i] = LLVector4(0,0,0,0);
-				}
-				else
-				{
-					clothing_weightsp[mMesh->mFaceVertexOffset + i] = (*(mMesh->getClothingWeights() + i));
-				}
+				v[0] = *(coords++); 
+				v[1] = *(coords++); 
+				v[2] = *(coords++);
+				v += vert_skip;
+
+				tc[0] = *(tex_coords++); 
+				tc[1] = *(tex_coords++);
+				tc += tc_skip;
+
+				n[0] = *(normals++); 
+				n[1] = *(normals++);
+				n[2] = *(normals++);
+				n += n_skip;
+
+				vw[0] = *(weights++);
+				vw += vw_skip;
+
+				cw[0] = *(cloth_weights++);
+				cw[1] = *(cloth_weights++);
+				cw[2] = *(cloth_weights++);
+				cw[3] = *(cloth_weights++);
+				cw += cw_skip;
 			}
+			while (++i < num_verts);
+
+			const U32 idx_count = mMesh->getNumFaces()*3;
 
-			for (S32 i = 0; i < mMesh->getNumFaces(); i++)
+			U16* __restrict idx = indicesp.get();
+			S32* __restrict src_idx = (S32*) mMesh->getFaces();
+
+			i = 0;
+
+			do
 			{
-				for (U32 j = 0; j < 3; j++)
-				{
-					U32 k = i*3+j+mMesh->mFaceIndexOffset;
-					indicesp[k] = mMesh->getFaces()[i][j] + mMesh->mFaceVertexOffset;
-				}
+				*(idx++) = *(src_idx++);
 			}
+			while (++i < idx_count);
 		}
 	}
 }
diff --git a/indra/newview/llvoavatar.cpp b/indra/newview/llvoavatar.cpp
index 43a9bd2b2f..33ea0199b6 100644
--- a/indra/newview/llvoavatar.cpp
+++ b/indra/newview/llvoavatar.cpp
@@ -3730,8 +3730,8 @@ U32 LLVOAvatar::renderSkinnedAttachments()
 										LLJoint* joint = getJoint(skin->mJointNames[i]);
 										if (joint)
 										{
-											mat[i] = skin->mInvBindMatrix[i];
-											mat[i] *= joint->getWorldMatrix();
+											mat[i*2+0] = skin->mInvBindMatrix[i];
+											mat[i*2+1] = joint->getWorldMatrix();
 										}
 									}
 									
@@ -3739,6 +3739,10 @@ U32 LLVOAvatar::renderSkinnedAttachments()
 										skin->mJointNames.size(),
 										FALSE,
 										(GLfloat*) mat[0].mMatrix);
+									gSkinnedObjectSimpleProgram.uniformMatrix4fv("matrixPalette[0]", 
+										skin->mJointNames.size(),
+										FALSE,
+										(GLfloat*) mat[0].mMatrix);
 
 									buff->setBuffer(data_mask);
 
@@ -3747,7 +3751,9 @@ U32 LLVOAvatar::renderSkinnedAttachments()
 									S32 offset = face->getIndicesStart();
 									U32 count = face->getIndicesCount();
 
-									buff->drawRange(LLRender::TRIANGLES, start, end, count, offset);
+									glPointSize(8.f);
+									buff->drawRange(LLRender::POINTS, start, end, count, offset);
+									glPointSize(1.f);
 								}
 							}
 						}